The Week in AI: February 16-22, 2026 (Week 8)

This week’s news: February 16-22, 2026 (Week 8)

The week was dominated by the AI Impact Summit in Delhi, where tech companies announced investments worth hundreds of billions of dollars and 88 countries signed the non-binding New Delhi Declaration. On the model front, Anthropic released Claude Sonnet 4.6, Google shipped Gemini 3.1 Pro and Alibaba launched the open model Qwen3.5. Nvidia signed a multi-year deal with Meta for millions of chips and, according to media reports, is preparing to invest $30 billion in OpenAI. Meanwhile, Anthropic faced mounting pressure from the Pentagon, and its new security tool sent cybersecurity stocks tumbling.

New Models

Anthropic Releases Claude Sonnet 4.6

On Tuesday, February 17, Anthropic unveiled Claude Sonnet 4.6, a model the company says delivers performance close to its flagship models at a mid-tier price. The improvements focus mainly on coding, computer use (where the model clicks and types in applications on its own) and working with long documents.

The model became the default for both free and paid users of claude.ai and, in beta, supports a context window of 1 million tokens, meaning the amount of text the model can “see” at once. According to Anthropic’s internal tests, developers in Claude Code preferred the new Sonnet over the earlier Opus 4.5 in 59% of comparisons. (Verdict)

Google Launches Gemini 3.1 Pro

On Thursday, February 19, Google released a preview of Gemini 3.1 Pro. The company describes it as an upgraded “core” for solving complex tasks, which also powers the Gemini 3 Deep Think mode introduced the previous week.

The model is available in the Gemini app, in NotebookLM for Pro and Ultra subscribers, and for developers via Vertex AI, Google AI Studio and Gemini CLI. Google claims it more than doubles the performance of Gemini 3 Pro on reasoning benchmarks. (Google Cloud Blog, TechRadar)

Alibaba Opens Up Qwen3.5

On Monday, February 16, China’s Alibaba introduced Qwen3.5, a model built for the “agentic era”, meaning AI that carries out multi-step tasks on its own. The model can read a phone or computer screen and click buttons or fill in forms by itself.

According to Reuters, it is 60% cheaper to run than its predecessor. The company released the flagship variant Qwen3.5-397B-A17B as an open-weights model, so anyone can download it and run it themselves. (Reuters via Yahoo Finance)

February 17 – Elon Musk’s xAI launched a public beta of Grok 4.20, in which four specialized agents work together; for now, only SuperGrok and X Premium+ subscribers have access. (NextBigFuture)
February 18 – Indian startup Sarvam AI unveiled two in-house language models at the summit, Sarvam-30B and Sarvam-105B, focused on Indian languages and so-called sovereign AI (developed and operated domestically). (YourStory)

Products and Agents

Claude Code Security and the Cybersecurity Stock Slide

On Friday, February 20, Anthropic launched a limited preview of Claude Code Security, a tool that scans source code for security vulnerabilities. It is offered to enterprise customers as part of the web version of Claude Code.

Investors responded with a sell-off: shares of companies such as CrowdStrike, Cloudflare, Okta and JFrog fell sharply during the day. Some analysts, however, called the reaction overblown, since similar tools already exist, such as Aardvark from OpenAI. (SiliconANGLE, heise)

Gemini Composes Music with Lyria 3

On Wednesday, February 18, Google added the Lyria 3 music model from DeepMind to the Gemini app. Users aged 18 and over can turn text or a photo into a 30-second track, complete with lyrics, vocals and cover art.

Every track carries an invisible SynthID watermark that identifies it as AI-generated. Google stresses that the tool is not meant to imitate existing artists. (Google Blog, 9to5Google)

The AI Impact Summit in Delhi

From February 16 to 21, the Indian capital hosted the India AI Impact Summit, the first major AI summit organized by a country of the Global South. The event was opened by Prime Minister Narendra Modi and drew dozens of heads of state as well as the leaders of the biggest AI companies. (Wikipedia)

The New Delhi Declaration

The summit closed on February 21 with the adoption of the New Delhi Declaration, backed by 88 countries and international organizations, including the US and China. The document calls for inclusive, “human-centric” AI development, but it is non-binding.

According to minister Ashwini Vaishnaw, the summit brought investment pledges of more than $250 billion. Fortune summed up the week as rich in money but poor in binding rules. (The Week, Fortune)

Investments Announced at the Summit

February 18 – Microsoft announced it will invest around $50 billion in AI across the Global South by 2030, covering infrastructure, education and multilingual models. (AI Business)
February 18 – Alphabet CEO Sundar Pichai unveiled America-India Connect, a project for new subsea cables linking India, the US and the Southern Hemisphere. It builds on a $15 billion investment in an AI hub in Visakhapatnam. (YourStory)
February 19 – OpenAI became the first customer of the Tata group’s HyperVault data centers, starting with 100 MW of capacity and an outlook of 1 GW. TCS will also roll out ChatGPT Enterprise to hundreds of thousands of employees. (TechCrunch)
February 19 – Mukesh Ambani announced that Reliance and Jio will invest 10 trillion rupees (about $110 billion) over seven years in sovereign computing capacity and data centers. (Bloomberg)
February 16-17 – Anthropic opened an office in Bengaluru (India is its second-largest market) and the next day announced a partnership with Infosys, which will integrate Claude models into its Topaz platform. (TechCrunch)

Summit Oddities

February 18 – The private Galgotias University was forced to leave the expo after passing off a Chinese Unitree Go2 robot dog as its own product. The university later apologized. (AP via WTOP)
February 19 – During a group photo, Modi asked the company leaders to join hands. Sam Altman (OpenAI) and Dario Amodei (Anthropic) instead just raised their fists, and the picture went viral as a symbol of the rivalry between the two companies. (Bloomberg)

Business and Investment

Nvidia to Supply Meta with Millions of Chips

On Tuesday, February 17, Nvidia and Meta announced a multi-year deal under which Meta will deploy millions of Blackwell and Rubin GPUs. It will also be the first large-scale deployment of standalone Grace CPUs without GPUs.

The deal also includes Spectrum-X networking switches. The companies did not disclose its value; media estimates put it in the tens of billions of dollars. (Yahoo Finance, TechSpot)

OpenAI Heads Toward a Record Funding Round

According to Bloomberg, on February 19 OpenAI was nearing the close of the first phase of a funding round that could bring in more than $100 billion. The round could value the company at over $850 billion.

At the same time, the Financial Times reported that Nvidia will put roughly $30 billion into OpenAI. This investment replaces last year’s non-binding pledge of up to $100 billion, which was to be paid out in gradual installments. (Bloomberg, DatacenterDynamics)

Regulation and Law

Pentagon Threatens to Blacklist Anthropic

According to reports from February 16, Pentagon officials were considering designating Anthropic a “supply chain risk”. Such a designation would prevent military contractors from using Claude models in their work for the Department of Defense.

The dispute concerns the rules Anthropic places on military use of its models, such as its refusal to allow mass surveillance of American citizens. Secretary Pete Hegseth ultimately announced the designation only after the week covered here had ended. (SiliconANGLE)

ByteDance Backs Down to Hollywood

On February 16, ByteDance, the owner of TikTok, pledged to strengthen safeguards in its Seedance 2.0 video generator. Before that, film studios led by Disney had sent it cease-and-desist letters over copyright infringement, as the tool readily generated well-known movie characters.

The company also suspended the ability to upload photos of real people to the tool. (JoBlo, TheWrap)